List your company

Question Answering on ∞-Bench Longbook QA English (test)

11.2F1 ScoreNestedKV
4.55446.27978.0059.7303May 26, 2026

Evaluation results

MethodLinks
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.25, Input cap=40k2026.0511.2——
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.75, Input cap=40k2026.0511.2——
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.25, Input cap=40k2026.0510.98——
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.25, Input cap=40k2026.0510.76——
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.50, Input cap=40k2026.0510.58——
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.50, Input cap=40k2026.0510.4——
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.75, Input cap=40k2026.0510.19——
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.25, Input cap=40k2026.0510.13——
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.25, Input cap=40k2026.0510——
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.50, Input cap=40k2026.059.63——
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.50, Input cap=40k2026.059.47——
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.50, Input cap=40k2026.059.06——
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.75, Input cap=40k2026.057.77——
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.75, Input cap=40k2026.057.3——
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.75, Input cap=40k2026.056.79——
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.25, Input cap=40k2026.055.4——
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.50, Input cap=40k2026.055.11——
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.75, Input cap=40k2026.054.81——
InfLLMContext Window=10242026.03—1,02416.27
InfLLMContext Window=20482026.03—2,04818.38
InfLLMContext Window=40962026.03—4,09622.09
TokenSelectContext Window=10242026.03—1,02413.71
TokenSelectContext Window=20482026.03—2,04816.93
TokenSelectContext Window=40962026.03—4,09618.44
UT-ACAMaximum Block Budget=8, Update=Sub.42026.03—89616.58
UT-ACAMaximum Block Budget=16, Update=Sub.82026.03—1,66418.72
UT-ACAMaximum Block Budget=32, Update=Sub.162026.03—3,71221.32